Ask Anna

Ask Anna landing and answer screens side by side
Team
1 PM, 5 Engineers, 3 Data Scientists & 7 Customer Success Managers
role
Lead Product designer

TIMELINE
Nov 2025 – Jul 2026
Outcome
‍

3h→20m

Faster analysis

Users significantly reduced the time needed to produce an intelligence analysis.
‍

25% ↑

Answer reuse
More analysts reused AI-generated answers, showing they were clear and actionable.

3x

source clicks
Citation engagement tripled. Analysts were not just reading answers, they were verifying them.

Overview

Seerist helps analysts, security teams, and government organizations understand global risks and make informed decisions. But analysts were overwhelmed by the amount of intelligence they had to review
The real challenge was reaching a clear answer fast enough to act on it
Ask Anna used AI to surface clear, actionable answers faster. But speed alone was not enough. Because these answers informed high-stakes decisions, analysts needed to understand and verify the information before they could trust it.
An answer that was incorrect, unsupported, or misunderstood could carry real consequences.
The Problem
Ask Anna looked like a conversational assistant, so users treated it like one.
They asked broad questions, expected the system to understand missing context, and assumed an unexpected answer meant the AI was wrong or broken.
In reality, Ask Anna behaved more like structured search. It worked best when users asked focused questions and provided enough context.
What users expected
AI Chat
  • Ask anything, in any way
  • System infers missing context
  • No right or wrong way to ask
≠
How Ask Anna actually worked
Structured intelligence search
  • Focused questions work best
  • Context must be provided by the user
  • Query quality shapes answer quality
"I assumed it worked like ChatGPT. When it didn't, I thought the AI was broken."
Intelligence Analyst — user research session
That mismatch created the wrong mental model, reduced trust, and caused some analysts to return to manual search.
The question I focused on
How might we design an AI experience that helps analysts understand how Ask Anna works, so they can use it confidently in a high stakes intelligence platform?

What Research Changed

I combined analyst interviews, user feedback, usability testing, and AI-assisted synthesis of user conversations.
Research showed that trust was not shaped by accuracy alone.
Users judged the product through the way answers were structured, whether evidence was visible, and how clearly the system communicated errors or limitations.
That shifted the work from making the experience feel more conversational to making it more accurately reflect the product's real behavior.
Three principles guided the design:
01
Match the interface to the product
The experience should feel like structured intelligence search, not an open-ended conversation.
02
Teach through interaction
Guidance should appear while users work rather than depend on onboarding they may forget.
03
Make confidence visible
Trust should come from clear structure, readable answers, visible evidence, and descriptive error states.
Each principle maps directly to a design decision.
Research
  • Users needed guidance before asking
  • Trust depended on transparency
  • Analysts needed to see when AI was uncertain
→
How Ask Anna actually works
  • Helps users ask better questions
  • Makes trust visible
  • Designed for uncertainty
Design
01 From chat to search
I replaced the continuous chat pattern with a clearer question-and-answer structure.
A chat thread with a follow up question, beside a single search field returning a structured answer
Shifting from open conversation to structured search set accurate expectations before analysts submitted a query.
02 Guidance that appears when it matters
User research showed analysts don’t want to learn or remember multiple tools, especially under time pressure.
Instead of relying on onboarding, I introduced examples, prompt suggestions, and contextual instructions to help analysts understand how much detail to provide
The old About paragraph beside the four step sequence that replaced it
Before: 1 long text block
Ask Anna landing with the how it works sequence closed, then opened to four numbered steps
After: Guidance when you need it, out of the way when you don’t
03Trust in the answer
I treated readability, hierarchy, citations, and errors as part of the trust model.
An answer with a key judgement, themed sections, and numbered citations listing each report and its date
Long responses were broken into clearer sections. Related information was grouped together. Citations became easier to find, and errors were made more explicit so users could tell the difference between missing information, an unsupported answer, and a system issue.
These changes made responses easier to scan, verify, and act on.
Constraints & Tradeoffs
The long-term vision included richer interactions and more flexible workflows, but technical constraints limited the first release.
For the MVP, I prioritized what would reduce friction and build trust.
Long-term vision
  • Flexible output formats (timeline, report)
  • Fit part of workflow: send answers to Report Builder and save to Content Folders
  • Preview citations within answers
  • Multi-turn contextual conversations
MVP: what we shipped
  • Clear question-and-answer model
  • Readable, structured responses
  • Visible evidence and citations
  • Persistent in-workflow guidance
  • Recoverable, explicit error states
hardest compromise
Asking users to adapt to limitations in the technology. My role was to reduce that burden through interaction patterns, microcopy, and familiar search behaviors.
Validation
The redesign was evaluated through moderated usability testing with analysts, focusing on whether they understood Ask Anna, could ask focused questions, noticed citations, and knew what to do when results didn’t meet expectations.
Key Finding
Language was part of the interaction design.
Small changes to prompts, labels, and examples improved query quality and reduced confusion about what the system could do.
Design
A broad question returns nothing, is narrowed, answered, then a citation is opened to check the source
Reflection
What I shaped
I helped define how AI should behave inside Seerist and shaped an interaction model that balanced user needs, technical limitations, business goals, and the level of trust required for intelligence work.

This project changed how I think about the details around an AI experience. What seemed like small interface decisions often had an outsized effect on how people understood and used the system.

I learned that designing AI is not just about the answer it produces. It is about designing the experience around that answer so people know how to ask, how to interpret the response, and when to trust it.

Rusul Kamal  © 2026
Email
Linkedin